Soniox TTS Real-Time v2

Soniox TTS Real-Time v2 is a high-performance text-to-speech model designed for low-latency, conversational AI applications. Released in August 2026, the model supports over 60 languages with native-speaker-quality output. It is built upon an in-house architecture that integrates a custom audio codec and an efficient inference engine to achieve sub-200ms latency, enabling audio generation to begin while text is still being streamed from a source like a Large Language Model (LLM).

A core capability of the v2 model is its precision with alphanumerics, specifically addressing common failures in synthetic speech such as the pronunciation of phone numbers, email addresses, verification codes, and technical identifiers. The system ensures hallucination-free output, strictly adhering to the input text while maintaining natural prosody. It also provides character-level timestamps, allowing developers to synchronize visual elements with audio and manage clean interruptions in voice agent workflows.

For expressive control, the model introduces audio tags, which allow users to directly influence emotion, delivery style, and vocal reactions. It also features high-fidelity voice cloning, where a cloned voice retains full support for multilingual speech and mid-sentence language switching. The engine is optimized to reduce silence between sentences and punctuation, preventing unnatural pauses during long-form generation or multi-turn dialogues.

The model is accessible via WebSocket API, supporting concurrent streams for multi-user environments. Its architecture is designed for regional deployment in the US, EU, and Japan to meet specific data residency and latency requirements for production-scale voice infrastructure.